Ling 3.2 Flash review: free with 1M context, but very slow

By LLMTest Lab · Nov 13, 2026 · 4 min read linghotfreebenchmarks
On this page

On this page

  1. What inclusionAI shipped
  2. How we tested
  3. Ling 3.2 Flash vs Ling 3.1 Flash
  4. Ling 3.2 Flash vs DeepSeek V4.1 Flash
  5. Who should pick it

Ling 3.2 Flash is a free model with a 1M-token window that answers about as well as its paid siblings, but it thinks for a long time first. In our 8-prompt developer test it beat DeepSeek V4.1 Flash on 2 prompts and lost 3, and it beat Ling 3.1 Flash on 1 and lost 2, with the rest tied. The price is $0, yet an average answer took 121 seconds against 11.6 seconds for DeepSeek, so it suits batch jobs and not anything a user is waiting on.

Ling 3.2 Flash (free) Ling 3.1 Flash DeepSeek V4.1 Flash
Input / output per 1M tokens $0 / $0 $0.045 / $0.135 $0.038 / $0.50
Context window 1,048,576 1,000,000 1,048,576
Input text, image text text
Avg latency in our test 121 to 151 s 56 s 11.6 s
Avg output tokens per answer 4,398 to 4,875 4,068 1,691
Avg cost per answer $0.00000 $0.00072 $0.00185

Prices, context sizes and modalities come from OpenRouter's model list. The Ling 3.2 Flash figures are a range because it ran in two separate head-to-heads. Everything else comes from our runs.

What inclusionAI shipped

OpenRouter lists Ling 3.2 Flash as a hybrid reasoning model from inclusionAI that takes text and image input, with a 1M-token context and up to 200 images per request. Output can run to 262,144 tokens, and reasoning is on by default but optional. It appeared on 2026-11-13 as a free endpoint, and the listing points to a Hugging Face repository. The previous model, Ling 3.1 Flash, is text-only and described as a 560B mixture-of-experts model with 25B active parameters. We found no official parameter count or benchmark table for 3.2 Flash, so we do not quote one.

How we tested

We wrote 8 prompts that look like work a developer hands a model: fixing a double-processing race in a Node job queue, a PostgreSQL query where a naive filter drops rows, extracting a German invoice to JSON without guessing, rewriting a release note under a word limit, a TypeScript string | undefined error, finding one retry line in 40 log lines, answering a benchmark question with no data in context, and planning tool calls under a refund policy. We ran inclusionai/ling-3.2-flash:free against inclusionai/ling-3.1-flash and deepseek/deepseek-v4.1-flash-1017 through OpenRouter, on 2026-11-13. Claude Sonnet 5.5 judged every prompt twice with the answer positions swapped, and a win counts only when both orders agree. Judge disagreements count as ties. We capped answers at 3,000 output tokens, and the harness retries empty answers. Total spend for this post was $0.30 of the $5 cap. Full outputs are in the saved runs vs-ling-3-1-flash and vs-deepseek-v4-1-flash.

Eight prompts is a small sample. Treat the splits below as a direction, not a ranking.

Ling 3.2 Flash vs Ling 3.1 Flash

3.2 won 1 prompt, 3.1 won 2, and 5 ended tied. The 3.2 win was the job-queue race, where it produced a complete FOR UPDATE SKIP LOCKED fix. The judge's note:

Response A gives a correct diagnosis and a correct fix. The explanation is two sentences, as requested. The CTE with FOR UPDATE SKIP LOCKED and an UPDATE ... RETURNING is the standard, correct approach.

The two losses were the SQL query and the TypeScript error. On SQL the judge reported the 3.2 answer as empty, so the model spent its budget reasoning and returned nothing visible:

Response A is empty, so it gives no query and no explanation and fails the request entirely.

On the TypeScript prompt the 3.2 answer was cut off before it gave the third fix or its pick. Both models were truncated on 2 of 8 prompts, so this is not a flaw unique to 3.2. It is the cost of models that reason at length under a fixed cap. If you raise the cap, you pay in time.

Ling 3.2 Flash vs DeepSeek V4.1 Flash

DeepSeek won 3 prompts, Ling won 2, and 3 were tied. DeepSeek took the invoice (it left the shipping quantity as null, as instructed, where Ling guessed 1), the SQL query, and the log lookup. Ling took the TypeScript error and the release-note rewrite. The judge on the invoice:

The key trap is the shipping line, which has no explicit quantity. The instruction says to use null for anything missing and not to guess. Response A sets qty to 1 for shipping, which is an inference and so a small fabrication.

On the TypeScript prompt, DeepSeek was the one that got cut off, and Ling's complete answer won:

Response A is clearly better. It explains the cause (host? makes c.host string | undefined) and gives three distinct, correct fixes: a non-null assertion, a ?? default, and a throw-based narrowing guard.

Ling's SQL answer was again empty in this run, after about 12,000 output tokens. Latency is where the two split most: DeepSeek averaged 11.6 seconds and 1,691 output tokens per answer, Ling 120.9 seconds and 4,875.

Who should pick it

Use Ling 3.2 Flash when the job is offline and the price is the point: overnight classification, bulk document triage with images, long-context lookups where you can wait. A free 1M-window model with image input is rare. Skip it for interactive features, for strict "do not guess" extraction unless you validate the output, and for anything where an empty answer breaks a pipeline; add a retry and a fallback model.

If you need speed on a budget, DeepSeek V4.1 Flash cost about $0.0019 per answer here and answered in seconds. For other no-cost options, see our ranking of free LLMs and the cheap model shootout. We will rerun this when inclusionAI publishes official benchmarks or the free endpoint's latency changes.

Related articles

Gemini 2.5 Flash vs Haiku 4.5 vs GPT-5 mini: GPT-6 Luna beats all 3
We ran Gemini 2.5 Flash, Claude Haiku 4.5 and GPT-5 mini against two newer budget models on 8 real tasks. GPT-6 Luna won at $0.19 per 1,000 answers.
Best free LLMs in 2026: where each hits a wall
Eight free LLMs worth actually using in 2026 — ranked by quality ceiling, real rate limits, and the exact point each stops being enough.